CN112052894B - Real-time monitoring and evaluating method for power cabin state - Google Patents
Real-time monitoring and evaluating method for power cabin state Download PDFInfo
- Publication number
- CN112052894B CN112052894B CN202010910987.6A CN202010910987A CN112052894B CN 112052894 B CN112052894 B CN 112052894B CN 202010910987 A CN202010910987 A CN 202010910987A CN 112052894 B CN112052894 B CN 112052894B
- Authority
- CN
- China
- Prior art keywords
- fault
- layer
- neural network
- circulation
- sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Probability & Statistics with Applications (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Testing Or Calibration Of Command Recording Devices (AREA)
Abstract
The invention discloses a real-time monitoring and evaluating method for the state of a power cabin, which comprises the steps of collecting various state data of the power cabin through a multi-path sensor; slicing the state data according to a time dimension; extracting spatial features of the data slice by using a convolutional neural network with a sensor data calibration module; further extracting a characteristic sequence representing time on the spatial characteristics by using a gated cyclic neural network with a fault category analysis module so as to obtain fault parameters; processing each data slice in sequence to realize the extraction of fault types and fault parameters; and finally, calculating the total evaluation score of the running state of the power cabin by using the fault parameters of the current state data for a period of time according to an entropy method. By using the method, the fault category and the fault parameters representing the fault severity can be obtained at one time by using a small amount of sensor data, the score representing the running state of the power cabin is calculated, and the monitoring and evaluation of the state of the power cabin are completed.
Description
Technical Field
The invention relates to the technical field of automatic detection, in particular to a real-time monitoring and evaluating method for the state of a power cabin.
Background
The power cabin is one of the most important core components of an automobile, mainly comprises an engine, a gearbox, a cooling system and the like, and is used for providing power for the whole automobile. Such important components naturally require more attention and protection, and therefore accurate fault diagnosis algorithms are essential. The problem is found before the vehicle is loaded, so that the vehicle with the manufacturing problem can be prevented; when a problem is found in a running vehicle, the personal safety of other parts of the system and a driver can be protected.
The existing mature method generally only judges the fault type of the power cabin, but has no proper judgment method for the fault severity, and the fault severity is a key factor influencing the treatment measures.
Disclosure of Invention
In view of the above, the invention provides a real-time monitoring and evaluating method for the state of a power cabin, which can obtain fault types and fault parameters representing the severity of faults at one time by using a small amount of sensor data, calculate scores representing the running state of the power cabin, complete monitoring and evaluation on the state of the power cabin, ensure normal running of the power cabin and enhance the safety of an automobile.
In order to solve the technical problem, the invention is realized as follows:
a real-time monitoring and evaluating method for the state of a power cabin comprises the following steps:
s1, acquiring various state data of the power cabin through a multi-path sensor; each state data is a one-dimensional time sequence; slicing the state data according to a time dimension;
s2, extracting spatial features of the data slices by using a convolutional neural network with a sensor data calibration module; the sensor data calibration module amplifies state data of the reaction fault by using global average pooling;
s3, a gated cyclic neural network with a fault category analysis module is used for further extracting a characteristic sequence of the characteristic time on the spatial characteristics so as to obtain fault parameters; the fault category analysis module extracts a first characteristic sequence from the gated cyclic neural network main channel, processes the first characteristic sequence to obtain a fault type, and simultaneously feeds the fault type back to the gated cyclic neural network main channel to continue processing to generate a second characteristic sequence so as to determine fault parameters describing the severity of the fault;
processing each data slice in sequence by adopting the mode of the steps S1 to S3 to realize the extraction of the fault type and the fault parameter;
and S4, calculating a total evaluation score of the running state of the power cabin by using the fault parameters of the current state data for a period of time according to an entropy method.
Preferably, the step S1 includes: splicing the multi-path state data into a two-dimensional matrix form, and dividing the multi-path state data into a plurality of matrix slices of M multiplied by N according to time; m represents that the matrix slice comprises M time sequences, and N is the total category number of the state data;
the step S2 includes:
s2-1, inputting the matrix slice into a first convolution module to obtain a first characteristic diagram;
s2-2, performing global average pooling on the first characteristic diagram according to the dimension of the state data type, obtaining N weight parameters within the range of [0,1] through a first full-connection network with an activation function being a sigmod function, and multiplying the weight parameters by the first characteristic diagram to obtain calibrated state data;
and S2-3, inputting the calibrated state data into a second convolution module to obtain a second characteristic diagram.
Preferably, the main channel of the gated recurrent neural network comprises at least 5 recurrent layers;
inputting a second feature map obtained by the convolutional neural network into a first circulation layer A of a main channel in the gated circulation neural network, and processing at least two circulation layers to obtain a first feature sequence; generating a circulating layer of the first characteristic sequence, namely a circulating layer In;
the first characteristic sequence is divided into two branches, one branch is input into a fully-connected network classifier serving as the fault category analysis module for fault type identification, and the other branch is input into a subsequent cycle layer of a main channel of a gated cyclic neural network;
the full-connection network classifier determines the fault type and further feeds the fault type back to the input of a circulation layer Fb of a main channel of the gated circulation neural network; at least one circulation layer is arranged between the circulation layer Fb and the circulation layer In at intervals, and at least one circulation layer is arranged between the circulation layer Fb and the last full-connection layer of the main channel of the gate control circulation neural network; the fully-connected network classifier comprises an input layer, at least 2 hidden layers and an output layer, wherein the output value of a neuron of the last hidden layer is fed back to a cycle layer Fb, and after the output value of the neuron of the last hidden layer is spliced with the output sequence of the cycle layer Fb into a new one-dimensional sequence according to the front-back sequence, the cycle layer Fb and all cycle layers behind the cycle layer Fb are sequentially processed to obtain a second characteristic sequence; a circulation layer is arranged between the circulation layer Fb and the full-connection layer at intervals, so that the circulation layer Fb is as close to an outlet of the gated circulation neural network as possible;
and the second characteristic sequence passes through the last full-connection layer of the main channel of the gated cyclic neural network to obtain the fault parameters.
Preferably, the main channel of the gated recurrent neural network comprises 5 layers of recurrent layers, named recurrent layers a, B, C, D, E; the first signature sequence is generated by a cyclic layer B; the circulating layer D is the circulating layer Fb.
Preferably, the step S4 includes the steps of:
s4-1, obtaining fault parameters of a period of time to form a matrix F, F = [ C = 1 ，C 2 ,…C G+1 ] T Total number of G fault types, C g A fault parameter vector for a G-th fault type, G =1,2, \8230;, G; c G+1 Corresponding to a normal state; each fault parameter vector comprises t fault parameters; when the same fault occurs for multiple times within the period of time, the average value of the fault parameters of the same fault for multiple times is taken and filled in the corresponding fault parameter position; normalizing according to fault parameter dimension, and setting x ij For the normalized fault parameter values of the ith row and the jth column in the F matrix, i =1,2, \8230, G +1, j =1,2, \8230, t;
s4-2, calculating the weight of each fault parameter by adopting an entropy method: firstly, calculating the proportion of each fault parameter of each fault category to the fault parameters
Secondly, calculating the entropy value of each fault parameter
Wherein s =1/ln (G + 1); then, the information entropy redundancy d of each fault parameter is calculated j ＝1-e j And using the ratio of the information entropy redundancy to the total fault parameter as the weight of the fault parameter
Step S4-3, according to
Calculating the score of each fault category, and finally summing the scores of all the fault categories to obtain the total evaluation score of the power cabin
Wherein, the G +1 fault types include: the method comprises the following steps of (1) in a normal state, in which an engine is not ignited, in which the engine is in poor idling, in which the engine is powerless, in which the engine is overheated, in which the engine oil pressure is abnormal, in which oil leaks, the fuel consumption is increased, and in which 4 faults exist among shafts of a gearbox, wherein the total number of the fault types is 12; the T fault parameters include the average amplitude change a caused by the fault, the change rate K of the amplitude change caused by the fault, the fault signal duration S, and the period T of the fault signal.
Has the beneficial effects that:
(1) The invention integrates fault type diagnosis and fault parameter acquisition, and is responsible for the fault diagnosis network, and the fault type and the fault parameters representing the fault severity are simultaneously obtained by adopting less sensor data. Meanwhile, the CNN-GRU network is improved, and a sensor data calibration module is added to the CNN network part and used for amplifying state data reflecting faults, so that the fault diagnosis is more accurate. In the GRU network part, the GRU main channel is adopted to realize the extraction of the fault parameters, the side branch of the type judgment is added on the GRU main channel, and the fault type judgment result is fed back to the main channel, so that the accuracy of the extraction of the fault parameters is improved.
(2) The invention also adopts the diagnosis results of fault category diagnosis and fault parameters to judge the severity of the fault, uses an entropy weight method to distribute the weight for various faults, uses the distributed weight to calculate the overall score of the system, and is used for monitoring and evaluating the state of the power cabin, so that the diagnosis and evaluation can be carried out on the whole, and the artificial subjective interference is reduced. And the judgment result of the fault severity also provides an important basis for the selection of subsequent treatment measures.
Drawings
FIG. 1 is a flow chart of a real-time monitoring and evaluating method for the state of a power compartment according to the present invention;
fig. 2 is a structural diagram of a fault diagnosis network based on an improved CNN-GRU network according to the present invention;
FIG. 3 is a block diagram of a sensor data calibration module;
FIG. 4 is a block diagram of a gated recurrent neural network with a fault class resolution module.
Detailed Description
The invention is described in detail below by way of example with reference to the accompanying drawings.
The invention provides a real-time monitoring and evaluating method for the state of a power cabin, which comprises three parts of fault type judgment, fault parameter acquisition and fault severity evaluation.
Firstly, as shown in fig. 1, the method integrates fault category diagnosis and fault parameter acquisition, and is responsible for a fault diagnosis network, and a fault category and a fault parameter representing the severity of a fault are simultaneously obtained by using less sensor data. The fault diagnosis network is realized by adopting a hybrid network of a convolutional neural network-gated cyclic neural network (CNN-GRU). The CNN network is mainly used in the image field and can be used for image classification, region separation, target identification and the like. The GRU network is mainly used in the field of natural language processing and can be used for sequence classification, numerical prediction, coding and decoding and the like. The CNN-GRU hybrid network combines the two to complete the image to text task, i.e. "speak with pictures". The problem from image to text is also one of the sequence to sequence problems, and input data is encoded by using CNN to obtain an abstract sequence, and the abstract sequence is decoded by using the GRU network. Therefore, the CNN-GRU hybrid network is more suitable for the characteristics with space-time characteristics.
Secondly, the invention improves the CNN-GRU network. The arrangement of each pixel data of the image is physically determined, and the arrangement of data sequences of different sensors is artificially designed. In order to reduce the influence caused by random arrangement, a sensor data calibration module is added in the CNN network part and is used for amplifying state data reflecting faults, so that the fault diagnosis is more accurate. In the GRU network part, the GRU main channel is adopted to realize the extraction of the fault parameters, the side branch of the type judgment is added on the GRU main channel, and the fault type judgment result is fed back to the main channel, so that the accuracy of the extraction of the fault parameters is improved.
In addition, the severity of the fault is characterized using the probability of the fault category and the fault parameters. The power compartment has a complex structure and a plurality of components, so that the types of faults are also large, and the diagnosis evaluation lacks integrity if the severity of each fault is only independently evaluated. Therefore, the invention provides that the weight distribution is carried out on various faults by using an entropy weight method, and the overall score of a system is calculated by using the distributed weight to be used for monitoring and evaluating the state of the power cabin.
Therefore, the method of the invention uses the improved CNN-GRU network to diagnose the fault, which is a data-driven supervised learning method, and the algorithm can automatically learn according to the labeled data, thereby avoiding the difficulty of feature selection. And the CNN-GRU network is added with a sensor data calibration module and a fault type analysis module, so that the fault type and fault parameter diagnosis accuracy can be improved. The invention also uses an entropy method to calculate the total state score, which is an objective weighting method and reduces man-made subjective interference.
Fig. 2 shows an original block diagram of the real-time monitoring and evaluating method for the state of the power cabin. As shown in the figure:
s1, acquiring various state data of the power cabin through a multi-path sensor; each state data is a one-dimensional time sequence; the state data is sliced in the time dimension.
In the step, the multi-channel state data are spliced into a two-dimensional matrix form and are divided into a plurality of MxN matrix slices according to time; m represents that the matrix slice comprises M time series, and N is the total category number of the state data. In the preferred embodiment of the invention, the power cabin is connected with a computer measurement and control system through various sensor interfaces, and parameters which can be directly read out on a computer comprise: the diesel engine comprises 8 state parameters including diesel flow, output rotating speed, output torque, engine oil pressure, engine oil temperature, water temperature, fan pump inlet pressure and fan pump outlet pressure. And if the number of sampling points is P, the sensor data after the same-frequency processing is P multiplied by 8. Generally, the time of one working cycle is taken as one slice, so that the data division and the network training are facilitated. The slice length M is selected according to the actual application context, in this example M =60. Dividing the sensor into [60 × 8] sized slices one by one in order can result in P-59 slices as input slices in the training sample.
Furthermore, common faults of the power pod are: the engine is not on fire, the engine is in poor idling, the engine is powerless, the engine is overheated, the engine oil pressure is abnormal, oil leakage, fuel consumption is increased, and faults occur among shafts of the gearbox (the gearbox has 7 gears in total, including a reverse gear, and has 4 fault states in total). The faults are K =11 fault states, and if a normal state is added, the power cabin has 12 states.
The fault parameters characterizing the severity of the fault are respectively: mean amplitude change a due to fault, rate of change K of amplitude change due to fault, fault signal duration S, and period T of fault signal.
S2, extracting spatial features of the data slice by using a convolutional neural network with a sensor data calibration module; the sensor data calibration module utilizes global average pooling to amplify the state data reflecting the fault.
As shown in fig. 2, the convolutional neural network includes a first convolution module, a sensor data calibration module, and a second convolution module. As shown in FIG. 3, the sensor data calibration module includes a pass-through channel and an active amplification channel. The through channel directly connects the output of the first convolution module with the input of the second convolution module; the active amplification path includes a global averaging pooling layer and a first fully connected network.
The step S2 specifically includes the following substeps:
and S2-1, inputting the matrix slice into a first convolution module to obtain a first characteristic diagram.
S2-2, performing global average pooling on the first characteristic diagram according to the dimension of the state data type, obtaining N weight parameters through a first full-connection network, and multiplying the weight parameters and the first characteristic diagram to obtain calibrated state data. The activation function of the first fully-connected network selects the sigmod function, resulting in data in the range between [0,1] as weights.
And S2-3, inputting the calibrated state data into a second convolution module to obtain a second characteristic diagram. The first feature map and the second feature map belong to spatial features.
And S2-4, taking down a next slice, and repeating the steps.
In a preferred embodiment, step S2 initializes 3 convolution kernels of 2 × 2 as the first convolution layer, slides through the slices of [60 × 8] sensor data (i.e., the state data), and fills the edges with 0 to ensure consistency of the sensor data format. If a feature is frequently activated in the convolutional layer, then the value calculated using the global average pooling is larger. Therefore, the convolution results of the first convolutional layer are subjected to global average pooling in the sensor dimension. If the result of the average pooling is directly used as the weight of the sensor, the network parameters can be changed drastically, so the weight is input into a 2-layer fully-connected network for further processing, and the sigmod function is selected as the activation function of the fully-connected network to obtain the sensor weight within the range of [0,1 ]; the weight is multiplied by the result output by the first convolutional layer and the result is fed to the next second convolutional layer, the convolutional kernel of the second convolutional layer is [2 x 2], and the number is 3. The above steps are repeated for each slice.
S3, further extracting a characteristic sequence representing time on the spatial characteristics by using a gated cyclic neural network with a fault category analysis module so as to obtain fault parameters; the fault category analysis module extracts a first characteristic sequence from the gated cyclic neural network main channel, processes the first characteristic sequence to obtain a fault type, and simultaneously feeds the fault type back to the gated cyclic neural network main channel to continue processing to generate a second characteristic sequence so as to determine fault parameters describing the severity of the fault.
The fault category analysis module adopts a full-connection network classifier. As shown in fig. 4, the main channel of the gated recurrent neural network of the present invention includes at least 5 recurrent layers.
Inputting a second feature map obtained by the convolutional neural network into a first circulation layer A of a main channel in the gated circulation neural network, and processing at least two circulation layers to obtain a first feature sequence; the circulation tank for generating the first characteristic sequence is the circulation layer In.
The first characteristic sequence is divided into two branches, one branch is input into a full-connection network classifier serving as the fault category analysis module to carry out fault type identification, and the other branch is input into a subsequent cycle layer of a gated cyclic neural network main channel.
And the full-connection network classifier determines the fault type and further feeds the fault type back to the input of the circulation layer Fb of the main channel of the gated recurrent neural network. At least one circulation layer is arranged between the circulation layer Fb and the circulation layer In at intervals, and at least one circulation layer is arranged between the circulation layer Fb and the last full-connection layer of the main channel of the gate control circulation neural network; the fully-connected network classifier comprises an input layer, at least 2 hidden layers and an output layer, wherein the output value of a neuron of the last hidden layer is fed back to the loop layer Fb, and after the output value of the neuron of the last hidden layer is spliced with the output sequence of the loop layer Fb, the loop layer Fb and all the subsequent loop layers are sequentially processed to obtain a second characteristic sequence; and a circulation layer is arranged between the circulation layer Fb and the full-connection layer in an interval mode, so that the circulation layer Fb is as close to the outlet of the gated recurrent neural network as possible.
And the second characteristic sequence passes through the last full-connection layer of the main channel of the gated cyclic neural network to obtain the fault parameters. Wherein the splicing may be a simple splicing of two sequences end to end.
In a preferred embodiment, the main channel of the gated recurrent neural network used in step S3 includes 5 recurrent layers, named recurrent layers a, B, C, D, E; a fully connected network classifier is also included. And modifying the spatial characteristic sequence obtained by the second convolutional layer into an input format of the cyclic network. The cycle layer a is set to contain 80 cycle cores, the cycle layer B is set to contain 100 GRUs, and the obtained result is recorded as a first signature sequence. Inputting the first characteristic sequence into a full-connection network classifier with a 4-layer structure, namely an input layer, a hidden layer 1, a hidden layer 2 and an output layer, wherein the number of neurons is 80, 50, 20 and 3 respectively, converting an output result into a classification probability through a softmax function, and taking the maximum probability as a final fault category. The first signature sequence is further processed by a loop layer C and a loop layer D, the number of GRUs being 50 and 40, respectively. The output of the loop layer D (length 40) is concatenated with the output of the fully-connected network classifier (length 20) to obtain a time series of length 60, which is denoted as a second time feature. And (4) passing the second time characteristic through a circulation layer E with 20 GRUs in the layer 1 and then through a full-connection network to obtain an output result, namely the fault parameter.
Processing each data slice in sequence by adopting the mode of the steps S1 to S3 to realize the extraction of the fault type and the fault parameter;
and S4, calculating a total evaluation score of the running state of the power cabin by using the fault parameters of the current state data for a period of time according to an entropy method.
The step S4 includes the steps of:
s4-1, obtaining fault parameters of a period of time to form a matrix F, F = [ C = 1 ，C 2 ,…C 12 ] T G is the total number of types of faults, and G =11 in this embodiment. C g A fault parameter vector for the g-th fault type, g =1,2, \ 8230, 12.C 12 Corresponding to the normal state. Each fault parameter vector has t =4 fault parameters, C g ＝[A g ,K g ,S g ,T g ]. When the same fault occurs for a plurality of times within the period of time, the average value of the fault parameters of the same fault for a plurality of times is filled in C g (ii) a Normalizing according to fault parameter dimension, and setting x ij For the normalized fault parameter values of the ith row and the jth column in the F matrix, i =1,2, \8230, G +1, j =1,2, \8230, t;
s4-2, calculating the weight of each fault parameter by adopting an entropy method: first, the proportion of each fault parameter of each fault category in the fault parameters is calculated and recorded as
Secondly, calculating the entropy value of each fault parameter and recording the entropy value as
Wherein s =1/ln (12); then, the information entropy redundancy d of each fault parameter is calculated j ＝1-e j And using the ratio of the information entropy redundancy to the total fault parameter as the weight of the fault parameter
Step S4-3, according to
Calculating the influence degree score of each fault category according to
And calculating the total evaluation score of the whole power cabin at the current moment.
The total score evaluation is mainly the total influence caused by various faults, so that the lower the total score is, the better the system operation state is indicated.
The training process of the network comprises the following steps: collecting sensor data containing faults in a test environment, making labels for each sample according to fault conditions, feeding the 8 paths of sensor data, 1 fault category label and 4 fault parameter labels into a network, and respectively calculating a cross entropy loss function of the fault categories and a mean square error loss function of the fault parameters, wherein a total loss function is the sum of the two multiplied by different weights. By utilizing an error back propagation algorithm, a root mean square transfer optimizer RMSprop is used for optimization, the initial learning rate is 0.001, the batch size is 64, the training times are 50 rounds, and the weight ratio of a cross entropy loss function to a mean square error loss function is set to be 1 in the first 20 rounds: 3, round 20 to round 40 are set to 3:1, last 10 rounds set to 1:1.
the above embodiments are merely illustrative of the design principles of the present invention, and the shapes of the components in the description may be different and the names are not limited. Therefore, a person skilled in the art of the present invention can modify or substitute the technical solutions described in the foregoing embodiments; such modifications and substitutions do not depart from the spirit and scope of the present invention.
Claims (6)
1. A real-time monitoring and evaluating method for the state of a power cabin is characterized by comprising the following steps:
s1, acquiring various state data of the power cabin through a multi-path sensor; each state data is a one-dimensional time sequence; slicing the state data according to a time dimension;
s2, extracting spatial features of the data slices by using a convolutional neural network with a sensor data calibration module; the sensor data calibration module amplifies state data of the reaction fault by using global average pooling;
s3, using a gated cyclic neural network with a fault category analysis module to extract a characteristic sequence representing time from the spatial characteristics so as to obtain fault parameters; the fault type analysis module extracts a first characteristic sequence from the gated cyclic neural network main channel, processes the first characteristic sequence to obtain a fault type, feeds the fault type back to the gated cyclic neural network main channel to be continuously processed to generate a second characteristic sequence, and further determines fault parameters describing the severity of the fault, wherein the first characteristic sequence and the second characteristic sequence are characteristic sequences representing time;
processing each data slice in sequence by adopting the mode of the steps S1 to S3 to realize the extraction of the fault type and the fault parameter;
and S4, calculating a total evaluation score of the running state of the power cabin by using the fault parameters of the current state data for a period of time according to an entropy method.
2. The method of claim 1, wherein the step S1 comprises: splicing the multi-path state data into a two-dimensional matrix form, and dividing the multi-path state data into a plurality of matrix slices of M multiplied by N according to time; m represents that the matrix slice comprises M time sequences, and N is the total category number of the state data;
the step S2 includes:
s2-1, inputting the matrix slice into a first convolution module to obtain a first characteristic diagram;
s2-2, performing global average pooling on the first characteristic diagram according to the dimension of the state data type, obtaining N weight parameters within the range of [0,1] through a first full-connection network with an activation function being a sigmod function, and multiplying the weight parameters by the first characteristic diagram to obtain calibrated state data;
and S2-3, inputting the calibrated state data into a second convolution module to obtain a second characteristic diagram.
3. The method of claim 1,
the main channel of the gated recurrent neural network at least comprises 5 recurrent layers;
inputting a second characteristic diagram obtained by the convolutional neural network into a first layer of circulation layer A of a main channel in the gated circulation neural network, and obtaining a first characteristic sequence through processing of at least two layers of circulation layers; the circulating layer for generating the first characteristic sequence is the circulating layer In;
the first characteristic sequence is divided into two branches, one branch is input into a fully-connected network classifier serving as the fault category analysis module for fault type identification, and the other branch is input into a subsequent cycle layer of a main channel of a gated cyclic neural network;
the full-connection network classifier determines the fault type and feeds the fault type back to the input of a circulation layer Fb of a main channel of the gated circulation neural network; at least one circulation layer is arranged between the circulation layer Fb and the circulation layer In at intervals, and at least one circulation layer is arranged between the circulation layer Fb and the last full-connection layer of the main channel of the gate control circulation neural network; the fully-connected network classifier comprises an input layer, at least 2 hidden layers and an output layer, wherein the output value of a neuron of the last hidden layer is fed back to a cycle layer Fb, and after the output value of the neuron of the last hidden layer is spliced with the output sequence of the cycle layer Fb into a new one-dimensional sequence according to the front-back sequence, the cycle layer Fb and all cycle layers behind the cycle layer Fb are sequentially processed to obtain a second characteristic sequence; a circulation layer is arranged between the circulation layer Fb and the full-connection layer at intervals, so that the circulation layer Fb is as close to an outlet of the gated circulation neural network as possible;
and the second characteristic sequence passes through the last full-connection layer of the main channel of the gated cyclic neural network to obtain the fault parameters.
4. The method of claim 3, wherein the main channel of the gated recurrent neural network comprises 5 layers of recurrent layers, named recurrent layers A, B, C, D, E; the first signature sequence is generated by a cyclic layer B; the circulation layer D is the circulation layer Fb.
5. The method of claim 1, wherein the step S4 comprises the steps of:
step S4-1, obtaining fault parameters of a period of time to form a matrix F, F = [ C = 1 ，C 2 ,…C G+1 ] T G total number of fault types, C g A fault parameter vector for a G-th fault type, G =1,2, \8230;, G; c G+1 Corresponding to a normal state; each fault parameter vector has t fault parameters; when the same fault occurs for multiple times within the period of time, the average value of the multiple fault parameters of the same fault is taken and filled in the corresponding fault parameter position; normalizing according to fault parameter dimension, and setting x ij For the normalized fault parameter values of the ith row and the jth column in the F matrix, i =1,2, \8230, G +1, j =1,2, \8230, t;
s4-2, calculating the weight of each fault parameter by adopting an entropy method: firstly, calculating the proportion of each fault parameter of each fault category to the fault category
Secondly, calculating the entropy value of each fault parameter
Wherein s =1/ln (G + 1); then, the information entropy redundancy d of each fault parameter is calculated j ＝1-e j And using the ratio of the information entropy redundancy to the total fault parameter as the weight of the fault parameter
Step S4-3, according to
Calculating the score of each fault category, and finally summing the scores of all fault categories to obtain the total evaluation score of the power cabin
6. The method of claim 5, wherein the G +1 fault types include: the method comprises the following steps of (1) in a normal state, in which an engine is not ignited, in which the engine is in poor idling, in which the engine is powerless, in which the engine is overheated, in which the engine oil pressure is abnormal, in which oil leaks, the fuel consumption is increased, and in which 4 faults exist among shafts of a gearbox, wherein the total number of the fault types is 12;
the T fault parameters comprise average amplitude change A caused by faults, change rate K of amplitude change caused by faults, fault signal duration S and fault signal period T.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010910987.6A CN112052894B (en) | 2020-09-02 | 2020-09-02 | Real-time monitoring and evaluating method for power cabin state |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010910987.6A CN112052894B (en) | 2020-09-02 | 2020-09-02 | Real-time monitoring and evaluating method for power cabin state |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112052894A CN112052894A (en) | 2020-12-08 |
| CN112052894B true CN112052894B (en) | 2023-02-07 |
Family
ID=73606602
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010910987.6A Active CN112052894B (en) | 2020-09-02 | 2020-09-02 | Real-time monitoring and evaluating method for power cabin state |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112052894B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112883656B (en) * | 2021-04-14 | 2022-11-01 | 北京理工大学 | Power cabin failure risk evaluation method based on normal operation data |
| CN113326853A (en) * | 2021-06-16 | 2021-08-31 | 西安隆基智能技术有限公司 | Neural network based process parameter analysis method and equipment and computer storage medium |
| CN114067829A (en) * | 2021-10-21 | 2022-02-18 | 佳源科技股份有限公司 | Reactor fault diagnosis method and device, computer equipment and storage medium |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3360081A1 (en) * | 2015-11-12 | 2018-08-15 | Google LLC | Convolutional gated recurrent neural networks |
| CN108569607B (en) * | 2018-06-22 | 2020-10-27 | 西安理工大学 | Elevator fault early warning method based on bidirectional gating cyclic neural network |
| CN110610158A (en) * | 2019-09-16 | 2019-12-24 | 南京师范大学 | Human body posture identification method and system based on convolution and gated cyclic neural network |
| CN111523081B (en) * | 2020-05-01 | 2023-09-12 | 西北工业大学 | Aeroengine fault diagnosis method based on enhanced gate control circulating neural network |
-
2020
- 2020-09-02 CN CN202010910987.6A patent/CN112052894B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN112052894A (en) | 2020-12-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112052894B (en) | Real-time monitoring and evaluating method for power cabin state | |
| CN112131760B (en) | CBAM model-based prediction method for residual life of aircraft engine | |
| CN109446187B (en) | Method for monitoring health state of complex equipment based on attention mechanism and neural network | |
| CN105354587B (en) | A kind of method for diagnosing faults of wind-driven generator group wheel box | |
| CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
| CN111967486A (en) | Complex equipment fault diagnosis method based on multi-sensor fusion | |
| CN110555474A (en) | photovoltaic panel fault detection method based on semi-supervised learning | |
| CN112200244A (en) | Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training | |
| CN106017876A (en) | Wheel set bearing fault diagnosis method based on equally-weighted local feature sparse filter network | |
| CN113743016B (en) | Engine residual life prediction method based on self-encoder and echo state network | |
| CN112541511A (en) | Multi-channel time series data fault diagnosis method based on convolutional neural network | |
| CN114544155A (en) | AUV propeller multi-information-source fusion fault diagnosis method and system based on deep learning | |
| CN113392773B (en) | Transformer fault diagnosis method of convolutional neural network based on map information fusion | |
| CN112541510A (en) | Intelligent fault diagnosis method based on multi-channel time series data | |
| CN112861071B (en) | High-speed rail traction system anomaly detection method based on depth self-coding | |
| CN116304819A (en) | Nuclear reactor operation condition judging method based on LeNet-5 algorithm | |
| CN114357372A (en) | Aircraft fault diagnosis model generation method based on multi-sensor data driving | |
| CN115641263A (en) | Single-power equipment infrared image super-resolution reconstruction method based on deep learning | |
| CN116738308A (en) | Wind turbine generator system fault diagnosis method and system based on WOA-KELM algorithm | |
| CN117041972A (en) | Channel-space-time attention self-coding based anomaly detection method for vehicle networking sensor | |
| CN114878427A (en) | Trident-net based method and system for simultaneously measuring soot multi-parameter fields in flame | |
| CN109034202A (en) | A kind of avionics system mode identification method of deepness belief network | |
| CN115526227A (en) | Helicopter flight data instantaneous anomaly detection method | |
| CN111652867A (en) | Heavy-environment-free plant growth method based on deep morphological belief network | |
| CN118468197B (en) | Multichannel feature fusion vehicle networking abnormality detection method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |